2  * Blktrace replay utility - Play traces back
18  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
70  * @devnm: 	Device name -- 'sd*'
120  * @tip:	Pointer to per-thread information this IO is associated with
140 static int cpus_to_use = -1;		// Number of CPUs to use
154 static int no_stalls = 0;		// Boolean: Disable pre-stalls
237  * min - Return minimum of two integers
245  * minl - Return minimum of two longs
253  * usage - Display usage string and version
257 	fprintf(stderr, "Usage: btreplay -- version %s\n%s",   in usage()
262  * is_send_done - Returns true if sender should quit early
263  * @tip: Per-thread information
267 	return signal_done || tip->send_done;  in is_send_done()
271  * is_reap_done - Returns true if reaper should quit early
272  * @tip: Per-thread information
276 	return signal_done || (tip->send_done && tip->naios_out == 0);  in is_reap_done()
280  * ts2ns - Convert timespec values to a nanosecond value
285 	return ((__u64)(ts->tv_sec) * NS_TICKS) + (__u64)(ts->tv_nsec);  in ts2ns()
289  * ts2ns - Convert timeval values to a nanosecond value
293 	return ((__u64)(tp->tv_sec)) + ((__u64)(tp->tv_usec) * (__u64)1000);  in tv2ns()
297  * touch_memory - Force physical memory to be allocating it
301  * as planned - we see Linux just use a single area to /read/ from
318  * buf_alloc - Returns a page-aligned buffer of the specified size
334  * gettime - Returns current time 
338 	static int use_clock_gettime = -1;		// Which clock to use  in gettime()
364  * setup_signal - Set up a signal handler for the specified signum
382  * __set_cv - Increments a variable under condition variable control.
400  * __wait_cv - Waits for a variable under cond var control to hit a value
465  * wait_iter_start - Wait for an iteration to start 
468  * non-zero, and then we decrement it and go on. 
476 	iter_start--;  in wait_iter_start()
481  * start_iter - Start an iteration at the replay thread level
499  * get_ncpus - Sets up the global 'ncpus' value
524 	ncpus = -1;  in get_ncpus()
538  * pin_to_cpu - Pin this thread to a specific CPU
549 	assert(0 <= tip->cpu && tip->cpu < ncpus);  in pin_to_cpu()
552 	CPU_SET_S(tip->cpu, size, cpus);  in pin_to_cpu()
557 	assert(tip->cpu == sched_getcpu());  in pin_to_cpu()
564 		fprintf(tip->vfp, "Pinned to CPU %02d ", tip->cpu);  in pin_to_cpu()
566 			fprintf(tip->vfp, "%1d", CPU_ISSET_S(i, size, now));  in pin_to_cpu()
567 		fprintf(tip->vfp, "\n");  in pin_to_cpu()
578  * add_input_dev - Add a device ('sd*') to the list of devices to handle
587 		if (strcmp(dip->devnm, devnm) == 0)  in add_input_dev()
592 	dip->devnm = strdup(devnm);  in add_input_dev()
593 	list_add_tail(&dip->head, &input_devs);  in add_input_dev()
597  * rem_input_dev - Remove resources associated with this device
601 	list_del(&dip->head);  in rem_input_dev()
602 	free(dip->devnm);  in rem_input_dev()
619 		if (strstr(ent->d_name, ".replay.") == NULL)  in find_input_devs()
622 		dsf = strdup(ent->d_name);  in find_input_devs()
640  * read_map_devs - Read in a set of device mapping from the provided file.
659 		mdp->from_dev = from_dev;  in read_map_devs()
660 		mdp->to_dev = to_dev;  in read_map_devs()
661 		list_add_tail(&mdp->head, &map_devs);  in read_map_devs()
668  * release_map_devs - Release resources associated with device mappings.
677 		list_del(&mdp->head);  in release_map_devs()
679 		free(mdp->from_dev);  in release_map_devs()
680 		free(mdp->to_dev);  in release_map_devs()
686  * map_dev - Return the mapped device for that specified
698 		if (strcmp(from_dev, mdp->from_dev) == 0)  in map_dev()
699 			return mdp->to_dev;  in map_dev()
712  * iocb_init - Initialize the fields of an IOCB
713  * @tip: Per-thread information
718 	iocbp->tip = tip;  in iocb_init()
719 	iocbp->nbytes = 0;  in iocb_init()
720 	iocbp->iocb.u.c.buf = NULL;  in iocb_init()
724  * iocb_setup - Set up an iocb with this AIOs information
733 	struct iocb *iop = &iocbp->iocb;  in iocb_setup()
739 	if (iocbp->nbytes) {  in iocb_setup()
740 		if (iocbp->nbytes >= n) {  in iocb_setup()
741 			buf = iop->u.c.buf;  in iocb_setup()
745 		assert(iop->u.c.buf);  in iocb_setup()
746 		free(iop->u.c.buf);  in iocb_setup()
750 	iocbp->nbytes = n;  in iocb_setup()
754 		io_prep_pread(iop, iocbp->tip->ofd, buf, n, off);  in iocb_setup()
757 		io_prep_pwrite(iop, iocbp->tip->ofd, buf, n, off);  in iocb_setup()
761 	iop->data = iocbp;  in iocb_setup()
766  * ==== PER-THREAD SET UP & TEAR DOWN =====================================
771  * tip_init - Per thread initialization function
777 	INIT_LIST_HEAD(&tip->free_iocbs);  in tip_init()
778 	INIT_LIST_HEAD(&tip->used_iocbs);  in tip_init()
780 	pthread_mutex_init(&tip->mutex, NULL);  in tip_init()
781 	pthread_cond_init(&tip->cond, NULL);  in tip_init()
783 	if (io_setup(naios, &tip->ctx)) {  in tip_init()
788 	tip->ofd = -1;  in tip_init()
789 	tip->naios_out = 0;  in tip_init()
790 	tip->send_done = tip->reap_done = 0;  in tip_init()
791 	tip->send_wait = tip->reap_wait = 0;  in tip_init()
793 	memset(&tip->sub_thread, 0, sizeof(tip->sub_thread));  in tip_init()
794 	memset(&tip->rec_thread, 0, sizeof(tip->rec_thread));  in tip_init()
800 		list_add_tail(&iocbp->head, &tip->free_iocbs);  in tip_init()
802 	tip->naios_free = naios;  in tip_init()
807 		sprintf(fn, "%s/%s.%s.%d.rep", idir, tip->devnm, ibase,   in tip_init()
808 			tip->cpu);  in tip_init()
809 		tip->vfp = fopen(fn, "w");  in tip_init()
810 		if (!tip->vfp) {  in tip_init()
815 		setlinebuf(tip->vfp);  in tip_init()
818 	if (pthread_create(&tip->sub_thread, NULL, replay_sub, tip)) {  in tip_init()
824 	if (pthread_create(&tip->rec_thread, NULL, replay_rec, tip)) {  in tip_init()
832  * tip_release - Release resources associated with this thread
838 	assert(tip->send_done);  in tip_release()
839 	assert(tip->reap_done);  in tip_release()
840 	assert(list_len(&tip->used_iocbs) == 0);  in tip_release()
841 	assert(tip->naios_free == naios);  in tip_release()
843 	if (pthread_join(tip->sub_thread, NULL)) {  in tip_release()
847 	if (pthread_join(tip->rec_thread, NULL)) {  in tip_release()
852 	io_destroy(tip->ctx);  in tip_release()
854 	list_splice(&tip->used_iocbs, &tip->free_iocbs);  in tip_release()
855 	list_for_each_safe(p, q, &tip->free_iocbs) {  in tip_release()
858 		list_del(&iocbp->head);  in tip_release()
859 		if (iocbp->nbytes)   in tip_release()
860 			free(iocbp->iocb.u.c.buf);  in tip_release()
864 	pthread_cond_destroy(&tip->cond);  in tip_release()
865 	pthread_mutex_destroy(&tip->mutex);  in tip_release()
869  * add_input_file - Allocate and initialize per-input file structure
877 	struct io_file_hdr hdr;  in add_input_file()  local
883 	memset(&hdr, 0, sizeof(hdr));  in add_input_file()
885 	tip->cpu = cpu % cpus_to_use;  in add_input_file()
886 	tip->iterations = def_iterations;  in add_input_file()
888 	tip->ifd = open(file_name, O_RDONLY);  in add_input_file()
889 	if (tip->ifd < 0) {  in add_input_file()
893 	if (fstat(tip->ifd, &buf) < 0) {  in add_input_file()
897 	if (buf.st_size < (off_t)sizeof(hdr)) {  in add_input_file()
903 	if (read(tip->ifd, &hdr, sizeof(hdr)) != sizeof(hdr)) {  in add_input_file()
908 	if (hdr.version != my_version) {  in add_input_file()
910 			(long long unsigned)hdr.version,  in add_input_file()
911 			(long long unsigned)hdr.genesis,  in add_input_file()
912 			(long long unsigned)hdr.nbunches,  in add_input_file()
913 			(long long unsigned)hdr.total_pkts);  in add_input_file()
916 			(long)hdr.version, (long)my_version);  in add_input_file()
920 	if (hdr.nbunches == 0) {  in add_input_file()
922 		close(tip->ifd);  in add_input_file()
927 	if (hdr.genesis < genesis) {  in add_input_file()
930 				du64_to_sec(hdr.genesis),  in add_input_file()
931 				du64_to_nsec(hdr.genesis));  in add_input_file()
932 		genesis = hdr.genesis;  in add_input_file()
935 	tip->devnm = strdup(devnm);  in add_input_file()
936 	tip->file_name = strdup(file_name);  in add_input_file()
938 	list_add_tail(&tip->head, &input_files);  in add_input_file()
942 			(long long)hdr.genesis);  in add_input_file()
946  * rem_input_file - Release resources associated with an input file
947  * @tip: Per-input file information
951 	list_del(&tip->head);  in rem_input_file()
955 	close(tip->ofd);  in rem_input_file()
956 	close(tip->ifd);  in rem_input_file()
957 	free(tip->file_name);  in rem_input_file()
958 	free(tip->devnm);  in rem_input_file()
963  * rem_input_files - Remove all input files
975  * __find_input_files - Find input files associated with this device (per cpu)
984 		sprintf(full_name, "%s/%s.%s.%d", idir, dip->devnm, ibase, cpu);  in __find_input_files()
988 		add_input_file(cpu, dip->devnm, full_name);  in __find_input_files()
993 		fatal(NULL, ERR_ARGS, "No traces found for %s\n", dip->devnm);  in __find_input_files()
1002  * find_input_files - Find input files for all devices
1020  * reap_wait_aios - Wait for and return number of outstanding AIOs
1029 		pthread_mutex_lock(&tip->mutex);  in reap_wait_aios()
1030 		while (tip->naios_out == 0) {  in reap_wait_aios()
1031 			tip->reap_wait = 1;  in reap_wait_aios()
1032 			if (pthread_cond_wait(&tip->cond, &tip->mutex)) {  in reap_wait_aios()
1038 		naios = tip->naios_out;  in reap_wait_aios()
1039 		pthread_mutex_unlock(&tip->mutex);  in reap_wait_aios()
1047  * reclaim_ios - Reclaim AIOs completed, recycle IOCBs
1048  * @tip: Per-thread information
1059 		ndone = io_getevents(tip->ctx, 1, naios_out, events, NULL);  in reclaim_ios()
1071 	pthread_mutex_lock(&tip->mutex);  in reclaim_ios()
1073 		struct iocb_pkt *iocbp = evp->data;  in reclaim_ios()
1075                 if (evp->res != iocbp->iocb.u.c.nbytes) {  in reclaim_ios()
1078                               (long)evp->res, (long)evp->res2,  in reclaim_ios()
1079                               (long)iocbp->iocb.u.c.offset / nb_sec,   in reclaim_ios()
1080 			      (long)iocbp->iocb.u.c.nbytes / nb_sec);  in reclaim_ios()
1084 		list_move_tail(&iocbp->head, &tip->free_iocbs);  in reclaim_ios()
1087 	tip->naios_free += ndone;  in reclaim_ios()
1088 	tip->naios_out -= ndone;  in reclaim_ios()
1089 	naios_out = minl(naios_out, tip->naios_out);  in reclaim_ios()
1091 	if (tip->send_wait) {  in reclaim_ios()
1092 		tip->send_wait = 0;  in reclaim_ios()
1093 		pthread_cond_signal(&tip->cond);  in reclaim_ios()
1095 	pthread_mutex_unlock(&tip->mutex);  in reclaim_ios()
1105  * replay_rec - Worker thread to reclaim AIOs
1116 	assert(tip->send_done);  in replay_rec()
1117 	tip->reap_done = 1;  in replay_rec()
1130  * next_bunch - Retrieve next bunch of AIOs to process
1131  * @tip: Per-thread information
1140 	result = read(tip->ifd, &bunch->hdr, sizeof(bunch->hdr));  in next_bunch()
1141 	if (result != sizeof(bunch->hdr)) {  in next_bunch()
1145 		fatal(tip->file_name, ERR_SYSCALL, "Short hdr(%ld)\n",   in next_bunch()
1149 	assert(bunch->hdr.npkts <= BT_MAX_PKTS);  in next_bunch()
1151 	count = bunch->hdr.npkts * sizeof(struct io_pkt);  in next_bunch()
1152 	result = read(tip->ifd, &bunch->pkts, count);  in next_bunch()
1154 		fatal(tip->file_name, ERR_SYSCALL, "Short pkts(%ld/%ld)\n",   in next_bunch()
1163  * nfree_current - Returns current number of AIOs that are free
1173 	pthread_mutex_lock(&tip->mutex);  in nfree_current()
1174 	while (!is_send_done(tip) && ((nfree = tip->naios_free) == 0)) {  in nfree_current()
1175 		tip->send_wait = 1;  in nfree_current()
1176 		if (pthread_cond_wait(&tip->cond, &tip->mutex)) {  in nfree_current()
1182 	pthread_mutex_unlock(&tip->mutex);  in nfree_current()
1188  * stall - Stall for the number of nanoseconds requested
1195 	long long dreal, tclock = gettime() - rgenesis;  in stall()
1200 		fprintf(tip->vfp, "   stall(%lld.%09lld, %lld.%09lld)\n",  in stall()
1205 		dreal = oclock - tclock;  in stall()
1210 			fprintf(tip->vfp, "++ stall(%lld.%09lld) ++\n",  in stall()
1218 		tclock = gettime() - rgenesis;  in stall()
1223  * iocbs_map - Map a set of AIOs onto a set of IOCBs
1224  * @tip: Per-thread information
1237 	pthread_mutex_lock(&tip->mutex);  in iocbs_map()
1238 	assert(ntodo <= list_len(&tip->free_iocbs));  in iocbs_map()
1240 		__u32 rw = pkt->rw;  in iocbs_map()
1243 		if (!pkt->rw && !write_enabled)  in iocbs_map()
1247 			fprintf(tip->vfp, "\t%10llu + %10llu %c%c\n",  in iocbs_map()
1248 				(unsigned long long)pkt->sector,   in iocbs_map()
1249 				(unsigned long long)pkt->nbytes / nb_sec,  in iocbs_map()
1251 				(rw == 1 && pkt->rw == 0) ? '!' : ' ');  in iocbs_map()
1253 		iocbp = list_entry(tip->free_iocbs.next, struct iocb_pkt, head);  in iocbs_map()
1254 		iocb_setup(iocbp, rw, pkt->nbytes, pkt->sector * nb_sec);  in iocbs_map()
1256 		list_move_tail(&iocbp->head, &tip->used_iocbs);  in iocbs_map()
1257 		list[i] = &iocbp->iocb;  in iocbs_map()
1260 	tip->naios_free -= ntodo;  in iocbs_map()
1261 	assert(tip->naios_free >= 0);  in iocbs_map()
1262 	pthread_mutex_unlock(&tip->mutex);  in iocbs_map()
1266  * process_bunch - Process a bunch of requests
1267  * @tip: Per-thread information
1273 	struct iocb *list[bunch->hdr.npkts];  in process_bunch()
1275 	assert(0 < bunch->hdr.npkts && bunch->hdr.npkts <= BT_MAX_PKTS);  in process_bunch()
1276 	while (!is_send_done(tip) && (i < bunch->hdr.npkts)) {  in process_bunch()
1278 		int ntodo = min(nfree_current(tip), bunch->hdr.npkts - i);  in process_bunch()
1281 		iocbs_map(tip, list, &bunch->pkts[i], ntodo);  in process_bunch()
1283 			stall(tip, bunch->hdr.time_stamp - genesis);  in process_bunch()
1287 				fprintf(tip->vfp, "submit(%d)\n", ntodo);  in process_bunch()
1288 			ndone = io_submit(tip->ctx, ntodo, list);  in process_bunch()
1292 					tip->cpu, ntodo, ndone,   in process_bunch()
1297 			pthread_mutex_lock(&tip->mutex);  in process_bunch()
1298 			tip->naios_out += ndone;  in process_bunch()
1299 			assert(tip->naios_out <= naios);  in process_bunch()
1300 			if (tip->reap_wait) {  in process_bunch()
1301 				tip->reap_wait = 0;  in process_bunch()
1302 				pthread_cond_signal(&tip->cond);  in process_bunch()
1304 			pthread_mutex_unlock(&tip->mutex);  in process_bunch()
1307 			assert(i <= bunch->hdr.npkts);  in process_bunch()
1313  * reset_input_file - Reset the input file for the next iteration
1320 	struct io_file_hdr hdr;  in reset_input_file()  local
1322 	lseek(tip->ifd, 0, 0);  in reset_input_file()
1324 	if (read(tip->ifd, &hdr, sizeof(hdr)) != sizeof(hdr)) {  in reset_input_file()
1325 		fatal(tip->file_name, ERR_ARGS, "Header reread failed\n");  in reset_input_file()
1331  * replay_sub - Worker thread to submit AIOs that are being replayed
1344 	mdev = map_dev(tip->devnm);  in replay_sub()
1358 	tip->ofd = open(path, O_RDWR | O_DIRECT | oflags);  in replay_sub()
1359 	if (tip->ofd < 0) {  in replay_sub()
1365 	while (!is_send_done(tip) && tip->iterations--) {  in replay_sub()
1368 			fprintf(tip->vfp, "\n=== %d ===\n", tip->iterations);  in replay_sub()
1374 	tip->send_done = 1;  in replay_sub()
1388         "\t[ -c <cpus> : --cpus=<cpus>           ] Default: 1\n"        \
1389         "\t[ -d <dir>  : --input-directory=<dir> ] Default: .\n"        \
1390 	"\t[ -F        : --find-records          ] Default: Off\n"	\
1391         "\t[ -h        : --help                  ] Default: Off\n"      \
1392         "\t[ -i <base> : --input-base=<base>     ] Default: replay\n"   \
1393         "\t[ -I <iters>: --iterations=<iters>    ] Default: 1\n"        \
1394         "\t[ -M <file> : --map-devs=<file>       ] Default: None\n"     \
1395         "\t[ -N        : --no-stalls             ] Default: Off\n"      \
1396         "\t[ -x        : --acc-factor            ] Default: 1\n"	\
1397         "\t[ -v        : --verbose               ] Default: Off\n"      \
1398         "\t[ -V        : --version               ] Default: Off\n"      \
1399         "\t[ -W        : --write-enable          ] Default: Off\n"      \
1412 		.name = "input-directory",
1418 		.name = "find-records",
1430 		.name = "input-base",
1442 		.name = "map-devs",
1448 		.name = "no-stalls",
1454 		.name = "acc-factor",
1472 		.name = "write-enable",
1494 	while ((c = getopt_long(argc, argv, S_OPTS, l_opts, NULL)) != -1) {  in handle_args()
1558 			fprintf(stderr, "btreplay -- version %s\n",   in handle_args()
1601  * set_signal_done - Signal handler, catches signals & sets signal_done
1609  * main -